Explicit Temporal and Spatial Modeling for Multivariate Time Series Anomaly Detection: A Survey

Abstract

Multivariate time series anomaly detection (MTSAD) plays a critical role in monitoring complex systems, where anomalies often arise from both abnormal temporal dynamics and irregular inter-variable interactions. Despite extensive research, existing methods vary widely in how temporal and spatial dependencies are modeled, making it difficult to systematically understand their design principles and limitations. In this survey, we review MTSAD methods from a unified spatial–temporal representation learning perspective. We analyze how temporal dependencies and inter-channel relationships are captured, and organize existing approaches into a coherent taxonomy based on implicit and explicit guidance mechanisms. Temporal modeling strategies are analyzed in terms of implicit dependency learning and explicitly guided temporal priors, while spatial modeling approaches are categorized into implicit and explicit guidance mechanisms, depending on whether inter-variable relationships are learned in a purely data-driven manner or guided by explicitly designed structural priors. Through this structured analysis, we highlight key design considerations related to modeling flexibility, robustness, and interpretability. This survey aims to clarify the evolving landscape of spatial–temporal anomaly detection and to provide insights for developing more reliable, interpretable, and scalable MTSAD frameworks.